Skip to content
AI-EscapePublic
forked from pingdotgg/t3code

About

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

 
 

Latest commit

 

History

12,800 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

T3 Code logo

T3 Code

The open-source control plane for coding agents.

This fork (AI-Escape, branch t4)

This is T3 Code v0.1.11 (Orchestrator V2) with the session tooling of t4 built in. t4 runs one isolated T3 server per client, and its dashboard, instructions and container layout live in that repo. Apart from this section, the README is upstream's. The full design and the reasons behind it are in docs/t4/DESIGN.md.

Keeping in sync: upstream (pingdotgg/t3code) release tags, not its main, merge into t4, and main is kept equal to t4. t4's image pins a commit of this branch (T3_FORK_REF). Fork code lives in apps/server/src/t4/ where it can; every edit to an upstream file is marked with a t4 fork: comment, so grep -rn "t4 fork:" lists the merge surface. The server reports its version as <upstream release>-t4.<commit>.

Added

  • Handoffs and an obligation ledger between threads. The new MCP tools t3_handoff (done / ask / fyi, explicit ledger effects), t3_owed and t3_decisions (the user's past answers to ask-user questions).
  • Release managers. One per project: routed to by t3_handoff, created on first use, pinned, always run on releaseManager.model, and never settle.
  • Linked issues (Linear). link_issue, unlink_issue and list_thread_issues, stored on the thread like linked PRs. The open thread that links an issue owns it, and launches refuse an owned issue unless forced.
  • Wakes. Scheduled tasks gain one-shot (at, in) and cron schedules, plus a threadId target for waking another thread, under t4's limits: 8 pending per thread, 30 days out at most, intervals of 5 minutes or more.
  • Stall recovery. A turn silent for 15 min is interrupted and restarted with a note; a second stall within 6 h is stopped and pushed to the operator.
  • Operator HTTP API under /api/t4 for the t4 dashboard and scripts:
    • reads: state, pending, archived;
    • writes: launch, resume-limited, threads/:id/{message,restore,delete}.
  • Its own state in <userdata>/t4.sqlite, with its own migrations. T3's statev2.sqlite is untouched. On first start it imports the old thread tools' JSON state once.

Changed

  • t3_thread_send: messages carry the sender's header. When a message closes an unanswered question or cancels a running tool call, it says so.
  • t3_thread_organize settle:
    • refused for release managers, and while the thread owes anything;
    • retires the worktree's dev stack;
    • a later message cancels a pending settle.
  • t3_thread_launch / create_threads:
    • a live-thread guard and a parent→child ledger entry with a spawn footer;
    • model presets and linked issues;
    • a launch that names no model gets the client's defaultModelSelection, not the launching thread's model.
  • t3_thread_list / t3_thread_read: each thread's ledger standing, release-manager flag and linked issues.
  • Worktree naming: T4_WORKTREE_NAMING=issueId gives short, flat <worktrees>/<issue id or slug> directories and keeps the full branch name.
  • Orchestration instructions: the Claude prompt append and Codex developer instructions are rewritten to teach this flow.
  • Claude adapter:
    • a --settings launch arg is merged under T3's own settings; upstream dropped it;
    • an ask-user question or approval whose tool call was aborted, for example by a steered message, is cancelled; upstream left it pending.
  • Codex: T4_CODEX_DEFAULT_REASONING_EFFORT sets the effort the model picker pre-selects; upstream uses the catalog default, low for some models.
  • Scheduled-task UI (web and mobile): shows the one-shot and cron schedules that agents create.

Configuration

Variable Purpose
T4_SERVER_CONFIG JSON file: release-manager title and model, worktree naming, launch presets, stall recovery, legacy state dir, dev-retire script glob
T4_WORKTREE_NAMING issueId (t4) or branch (upstream)
T4_CODEX_DEFAULT_REASONING_EFFORT Codex effort the picker pre-selects
T4_LABEL, T4_UI_URL Client name and dashboard link in stall pushes

Developing

Tests run in the Node 26 build container: ./.t4-test.sh <paths> (the helper is local and gitignored; it installs the pnpm packageManager pins, since Node 26 has no corepack). Typecheck everything from the root: ./.t4-test.sh --cmd "cd /src && pnpm run typecheck". Format with pnpm exec vp fmt <files> before committing. Browser tests need Chromium, which that container lacks.


Download · iOS · Android · Web · Docs · Discord

T3 Code on desktop with a finished Claude thread, next to the phone app asking to approve a command from Codex

T3 Code runs the coding agents on your machine and gives them one fast interface. Start a thread at your desk, then approve the next step from your phone.

It works with Claude Code, Codex, Cursor, OpenCode, Grok Build, Antigravity, Kiro, Pi, Muse Code, fx, and any ACP agent. If an agent is set up on your computer, T3 Code can control it.

What you get

  • Your subscriptions. T3 Code does not resell tokens. It drives the agents you are already signed in to.
  • Every agent in one place. Change models in the middle of a thread. Let Claude hand a task to Codex.
  • Desktop, web, and phone. One server, and every client shows the same threads. There are apps for iOS and Android.
  • Remote ready. Run agents on any machine. Connect over your local network, Tailscale, or T3 Connect.
  • Git built in. Each thread can work in its own worktree. Review the diff, then commit, push, and open a pull request with one button.
  • Fast. Performance is a feature here. We measure it, and we treat a dropped frame as a bug.
  • Open. MIT licensed. If you don't like something, fork it.

Install

Desktop app

Download T3 Code for macOS, Windows, or Linux. You can also use a package manager:

Platform Command
macOS brew install --cask t3-code
Windows winget install T3Tools.T3Code or scoop install extras/t3code
Arch Linux yay -S t3code-bin
Nix nix profile add github:NixOS/nixpkgs/nixos-unstable#t3code
Debian, Ubuntu .deb from Releases

Command line

This installs the t3 command. Run t3 to start the server and open the web app.

curl -fsSL https://t3.codes/install.sh | sh

On Windows, in PowerShell:

irm https://t3.codes/install.ps1 | iex

To try it once without an install, run npx t3@latest.

Phone

Get the app from the App Store or Google Play. Then follow remote access to connect it to your computer.

Sign in to an agent

T3 Code needs one agent that is installed and signed in. For example, install Claude Code and run claude auth login, or install Codex CLI and run codex login.

Provider setup has the steps for each agent. Install T3 Code covers nightly builds, AppImage, WSL, and NixOS.

"Wait, what are you selling me?"

Nothing. We built T3 Code because we wanted the best possible development experience with agents. We were inspired by existing solutions like the Codex desktop app, Conductor, Claude Desktop and Cursor Glass, but none met our bar.

We wanted something performant, remote-ready, and truly open. If we ever go the wrong direction, we want you to have everything you need to fork and build the tool that you want.

Documentation

All docs has the full list.

Contributing

We have limited review capacity. Small, focused fixes are the most likely to be merged. Features need an approved discussion first.

License

MIT

About

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages